CMSC 471 Fall 2011 Class #16 Tuesday, October 25, 2011 Planning Professor Marie desJardins, mariedj@cs.umbc.edu

Planning Chapter 10.1-10.2, 10.4.2-10.4.4 and Chuck Dyer Some material adopted from notes by Andreas Geyer-Schulz and Chuck Dyer

Today’s Class What is planning? Approaches to planning GPS / STRIPS Situation calculus formalism [revisited] Partial-order planning

Planning Problem Find a sequence of actions that achieves a given goal when executed from a given initial world state. That is, given a set of operator descriptions (defining the possible primitive actions by the agent), an initial state description, and a goal state description or predicate, compute a plan, which is a sequence of operator instances, such that executing them in the initial state will change the world to a state satisfying the goal-state description. Goals are usually specified as a conjunction of goals to be achieved

Planning vs. Problem Solving Planning and problem solving methods can often solve the same sorts of problems Planning is more powerful because of the representations and methods used States, goals, and actions are decomposed into sets of sentences (usually in first-order logic) Search often proceeds through plan space rather than state space (though there are also state-space planners) Subgoals can be planned independently, reducing the complexity of the planning problem

Typical Assumptions Atomic time: Each action is indivisible No concurrent actions are allowed (though actions do not need to be ordered with respect to each other in the plan) Deterministic actions: The result of actions are completely determined—there is no uncertainty in their effects Agent is the sole cause of change in the world Agent is omniscient: Has complete knowledge of the state of the world Closed world assumption: everything known to be true in the world is included in the state description. Anything not listed is false.

Blocks World The blocks world is a micro-world that consists of a table, a set of blocks and a robot hand. Some domain constraints: Only one block can be on another block Any number of blocks can be on the table The hand can only hold one block Typical representation: ontable(a) ontable(c) on(b,a) handempty clear(b) clear(c) A B C TABLE

Major Approaches GPS / STRIPS Situation calculus Partial order planning Hierarchical decomposition (HTN planning) Planning with constraints (SATplan, Graphplan) Reactive planning

General Problem Solver The General Problem Solver (GPS) system was an early planner (Newell, Shaw, and Simon) GPS generated actions that reduced the difference between some state and a goal state GPS used Means-Ends Analysis Compare what is given or known with what is desired and select a reasonable thing to do next Use a table of differences to identify procedures to reduce types of differences GPS was a state space planner: it operated in the domain of state space problems specified by an initial state, some goal states, and a set of operations

Situation Calculus Planning Intuition: Represent the planning problem using first-order logic Situation calculus lets us reason about changes in the world Use theorem proving to “prove” that a particular sequence of actions, when applied to the situation characterizing the world state, will lead to a desired result

Situation Calculus Planning, cont. Initial state: a logical sentence about (situation) S0 At(Home, S0)  Have(Milk, S0)   Have(Bananas, S0)   Have(Drill, S0) Goal state: (s) At(Home,s)  Have(Milk,s)  Have(Bananas,s)  Have(Drill,s) Operators are descriptions of how the world changes as a result of the agent’s actions: (a,s) Have(Milk,Result(a,s))  ((a=Buy(Milk)  At(Grocery,s))  (Have(Milk, s)  a  Drop(Milk))) Result(a,s) names the situation resulting from executing action a in situation s. Action sequences are also useful: Result'(l,s) is the result of executing the list of actions (l) starting in s: (s) Result'([],s) = s (a,p,s) Result'([a|p]s) = Result'(p,Result(a,s))

Situation Calculus, cont. A solution is a plan that when applied to the initial state yields a situation satisfying the goal query: At(Home, Result'(p,S0))  Have(Milk, Result'(p,S0))  Have(Bananas, Result'(p,S0))  Have(Drill, Result'(p,S0)) Thus we would expect a plan (i.e., variable assignment through unification) such as: p = [Go(Grocery), Buy(Milk), Buy(Bananas), Go(HardwareStore), Buy(Drill), Go(Home)]

Situation Calculus: Blocks World Here’s an example of a situation calculus rule for the blocks world: Clear (X, Result(A,S))  [Clear (X, S)  ((A=Stack(Y,X)  A=Pickup(X))  (A=Stack(Y,X)  (holding(Y,S))  (A=Pickup(X)  (handempty(S)  ontable(X,S)  clear(X,S))))]  [A=Stack(X,Y)  holding(X,S)  clear(Y,S)]  [A=Unstack(Y,X)  on(Y,X,S)  clear(Y,S)  handempty(S)]  [A=Putdown(X)  holding(X,S)] English translation: A block is clear if (a) in the previous state it was clear and we didn’t pick it up or stack something on it successfully, or (b) we stacked it on something else successfully, or (c) something was on it that we unstacked successfully, or (d) we were holding it and we put it down. Whew!!! There’s gotta be a better way!

Situation Calculus Planning: Analysis This is fine in theory, but remember that problem solving (search) is exponential in the worst case Also, resolution theorem proving only finds a proof (plan), not necessarily a good plan So we restrict the language and use a special-purpose algorithm (a planner) rather than general theorem prover

Basic Representations for Planning Classic approach first used in the STRIPS planner circa 1970 States represented as a conjunction of ground literals at(Home)  have(Milk)  have(bananas) ... Goals are conjunctions of literals, but may have variables which are assumed to be existentially quantified at(?x)  have(Milk)  have(bananas) ... Do not need to fully specify state Non-specified either don’t-care or assumed false Represent many cases in small storage Often only represent changes in state rather than entire situation Unlike theorem prover, not seeking whether the goal is true, but is there a sequence of actions to attain it

Operator/Action Representation Operators contain three components: Action description Precondition - conjunction of positive literals Effect - conjunction of positive or negative literals which describe how situation changes when operator is applied Example: Op[Action: Go(there), Precond: At(here)  Path(here,there), Effect: At(there)  At(here)] All variables are universally quantified Situation variables are implicit preconditions must be true in the state immediately before operator is applied; effects are true immediately after At(here) ,Path(here,there) Go(there) At(there) , At(here)

Blocks World Operators Here are the classic basic operations for the blocks world: stack(X,Y): put block X on block Y unstack(X,Y): remove block X from block Y pickup(X): pickup block X putdown(X): put block X on the table Each will be represented by a list of preconditions a list of new facts to be added (add-effects) a list of facts to be removed (delete-effects) optionally, a set of (simple) variable constraints For example: preconditions(stack(X,Y), [holding(X), clear(Y)]) deletes(stack(X,Y), [holding(X), clear(Y)]). adds(stack(X,Y), [handempty, on(X,Y), clear(X)]) constraints(stack(X,Y), [XY, Ytable, Xtable])

Blocks World Operators II operator(stack(X,Y), Precond [holding(X), clear(Y)], Add [handempty, on(X,Y), clear(X)], Delete [holding(X), clear(Y)], Constr [XY, Ytable, Xtable]). operator(pickup(X), [ontable(X), clear(X), handempty], [holding(X)], [Xtable]). operator(unstack(X,Y), [on(X,Y), clear(X), handempty], [holding(X), clear(Y)], [handempty, clear(X), on(X,Y)], [XY, Ytable, Xtable]). operator(putdown(X), [holding(X)], [ontable(X), handempty, clear(X)], [Xtable]).

STRIPS Planning STRIPS maintains two additional data structures: State List - all currently true predicates. Goal Stack - a push down stack of goals to be solved, with current goal on top of stack. If current goal is not satisfied by present state, examine add lists of operators, and push operator and preconditions list on stack. (Subgoals) When a current goal is satisfied, POP it from stack. When an operator is on top stack, record the application of that operator on the plan sequence and use the operator’s add and delete lists to update the current state.

Typical Blocks World Planning Problem Initial state: clear(a) clear(b) clear(c) ontable(a) ontable(b) ontable(c) handempty Goal: on(b,c) on(a,b) A B C A plan: pickup(b) stack(b,c) pickup(a) stack(a,b) A B C

Another Blocks World Planning Problem Initial state: clear(a) clear(b) clear(c) ontable(a) ontable(b) ontable(c) handempty Goal: on(a,b) on(b,c) A B C A plan: pickup(a) stack(a,b) unstack(a,b) putdown(a) pickup(b) stack(b,c) A B C

Goal Interactions A B C A B C Goal state Initial state Simple planning algorithms assume that the goals to be achieved are independent Each can be solved separately and then the solutions concatenated This planning problem, called the “Sussman Anomaly,” is the classic example of the goal interaction problem: Solving on(A,B) first (by doing unstack(C,A), stack(A,B) will be undone when solving the second goal on(B,C) (by doing unstack(A,B), stack(B,C)). Solving on(B,C) first will be undone when solving on(A,B) Classic STRIPS could not handle this, although minor modifications can get it to do simple cases A B C A B C Goal state Initial state

Sussman Anomaly Goal state A B C A B C Initial state Achieve on(a,b) via stack(a,b) with preconds: [holding(a),clear(b)] |Achieve holding(a) via pickup(a) with preconds: [ontable(a),clear(a),handempty] ||Achieve clear(a) via unstack(_1584,a) with preconds: [on(_1584,a),clear(_1584),handempty] ||Applying unstack(c,a) ||Achieve handempty via putdown(_2691) with preconds: [holding(_2691)] ||Applying putdown(c) |Applying pickup(a) Applying stack(a,b) Achieve on(b,c) via stack(b,c) with preconds: [holding(b),clear(c)] |Achieve holding(b) via pickup(b) with preconds: [ontable(b),clear(b),handempty] ||Achieve clear(b) via unstack(_5625,b) with preconds: [on(_5625,b),clear(_5625),handempty] ||Applying unstack(a,b) ||Achieve handempty via putdown(_6648) with preconds: [holding(_6648)] ||Applying putdown(a) |Applying pickup(b) Applying stack(b,c) From [clear(b),clear(c),ontable(a),ontable(b),on(c,a),handempty] To [on(a,b),on(b,c),ontable(c)] Do: unstack(c,a) putdown(c) pickup(a) stack(a,b) unstack(a,b) putdown(a) pickup(b) stack(b,c) Goal state A B C A B C Initial state

State-Space Planning We initially have a space of situations (where you are, what you have, etc.) The plan is a solution found by “searching” through the situations to get to the goal A progression planner searches forward from initial state to goal state A regression planner searches backward from the goal This works if operators have enough information to go both ways Ideally this leads to reduced branching: the planner is only considering things that are relevant to the goal

Planning Heuristics Just as with search, we need an admissible heuristic that we can apply to planning states Estimate of the distance (number of actions) to the goal Planning typically uses relaxation to create heuristics Ignore all or selected preconditions Ignore delete lists (movement towards goal is never undone) Use state abstraction (group together “similar” states and treat them as though they are identical) – e.g., ignore fluents Assume subgoal independence (use max cost; or if subgoals actually are independent, can sum the costs) Use pattern databases to store exact solution costs of recurring subproblems

Plan-Space Planning An alternative is to search through the space of plans, rather than situations. Start from a partial plan which is expanded and refined until a complete plan that solves the problem is generated. Refinement operators add constraints to the partial plan and modification operators for other changes. We can still use STRIPS-style operators: Op(ACTION: RightShoe, PRECOND: RightSockOn, EFFECT: RightShoeOn) Op(ACTION: RightSock, EFFECT: RightSockOn) Op(ACTION: LeftShoe, PRECOND: LeftSockOn, EFFECT: LeftShoeOn) Op(ACTION: LeftSock, EFFECT: leftSockOn) could result in a partial plan of [RightShoe, LeftShoe]

Partial-Order Planning A linear planner builds a plan as a totally ordered sequence of plan steps A non-linear planner (aka partial-order planner) builds up a plan as a set of steps with some temporal constraints constraints of the form S1<S2 if step S1 must comes before S2. One refines a partially ordered plan (POP) by either: adding a new plan step, or adding a new constraint to the steps already in the plan. A POP can be linearized (converted to a totally ordered plan) by topological sorting

Least Commitment Non-linear planners embody the principle of least commitment only choose actions, orderings, and variable bindings that are absolutely necessary, leaving other decisions till later avoids early commitment to decisions that don’t really matter A linear planner always chooses to add a plan step in a particular place in the sequence A non-linear planner chooses to add a step and possibly some temporal constraints

Non-Linear Plan A non-linear plan consists of (1) A set of steps {S1, S2, S3, S4…} Each step has an operator description, preconditions and post-conditions (2) A set of causal links { … (Si,C,Sj) …} Meaning a purpose of step Si is to achieve precondition C of step Sj (3) A set of ordering constraints { … Si<Sj … } if step Si must come before step Sj A non-linear plan is complete iff Every step mentioned in (2) and (3) is in (1) If Sj has prerequisite C, then there exists a causal link in (2) of the form (Si,C,Sj) for some Si If (Si,C,Sj) is in (2) and step Sk is in (1), and Sk threatens (Si,C,Sj) (makes C false), then (3) contains either Sk<Si or Sj<Sk

The Initial Plan Every plan starts the same way Initial State S1:Start S2:Finish Initial State Goal State

RightShoeOn ^ LeftShoeOn Trivial Example Operators: Op(ACTION: RightShoe, PRECOND: RightSockOn, EFFECT: RightShoeOn) Op(ACTION: RightSock, EFFECT: RightSockOn) Op(ACTION: LeftShoe, PRECOND: LeftSockOn, EFFECT: LeftShoeOn) Op(ACTION: LeftSock, EFFECT: leftSockOn) S1:Start S2:Finish RightShoeOn ^ LeftShoeOn Steps: {S1:[Op(Action:Start)], S2:[Op(Action:Finish, Pre: RightShoeOn^LeftShoeOn)]} Links: {} Orderings: {S1<S2}

Solution Start Left Sock Right Sock Left Shoe Right Shoe Finish

POP Constraints and Search Heuristics Only add steps that achieve a currently unachieved precondition Use a least-commitment approach: Don’t order steps unless they need to be ordered Honor causal links S1  S2 that protect a condition c: Never add an intervening step S3 that violates c If a parallel action threatens c (i.e., has the effect of negating or clobbering c), resolve that threat by adding ordering links: Order S3 before S1 (demotion) Order S3 after S2 (promotion) c

Partial-Order Planning Example Goal: Have milk, bananas, and a drill

Resolving Threats Threat Demotion Promotion

Real-World Planning Domains Real-world domains are complex and don’t satisfy the assumptions of STRIPS or partial-order planning methods Some of the characteristics we may need to deal with: Modeling and reasoning about resources Representing and reasoning about time Planning at different levels of abstractions Conditional outcomes of actions Uncertain outcomes of actions Exogenous events Incremental plan development Dynamic real-time replanning } Scheduling } Planning under uncertainty } HTN planning

Hierarchical Decomposition Hierarchical decomposition, or hierarchical task network (HTN) planning, uses abstract operators to incrementally decompose a planning problem from a high-level goal statement to a primitive plan network Primitive operators represent actions that are executable, and can appear in the final plan Non-primitive operators represent goals (equivalently, abstract actions) that require further decomposition (or operationalization) to be executed There is no “right” set of primitive actions: One agent’s goals are another agent’s actions!

HTN Operator: Example OPERATOR decompose PURPOSE: Construction CONSTRAINTS: Length (Frame) <= Length (Foundation), Strength (Foundation) > Wt(Frame) + Wt(Roof) + Wt(Walls) + Wt(Interior) + Wt(Contents) PLOT: Build (Foundation) Build (Frame) PARALLEL Build (Roof) Build (Walls) END PARALLEL Build (Interior)

HTN Planning: Example

HTN Operator Representation Russell & Norvig explicitly represent causal links; these can also be computed dynamically by using a model of preconditions and effects Dynamically computing causal links means that actions from one operator can safely be interleaved with other operators, and subactions can safely be removed or replaced during plan repair Russell & Norvig’s representation only includes variable bindings, but more generally we can introduce a wide array of variable constraints

Truth Criterion Determining whether a formula is true at a particular point in a partially ordered plan is, in the general case, NP-hard Intuition: there are exponentially many ways to linearize a partially ordered plan In the worst case, if there are N actions unordered with respect to each other, there are N! linearizations Ensuring soundness of the truth criterion requires checking the formula under all possible linearizations Use heuristic methods instead to make planning feasible Check later to be sure no constraints have been violated

Truth Criterion in HTN Planners Heuristic: prove that there is one possible ordering of the actions that makes the formula true – but don’t insert ordering links to enforce that order Such a proof is efficient Suppose you have an action A1 with a precondition P Find an action A2 that achieves P (A2 could be initial world state) Make sure there is no action necessarily between A2 and A1 that negates P Applying this heuristic for all preconditions in the plan can result in infeasible plans

Increasing Expressivity Conditional effects Instead of having different operators for different conditions, use a single operator with conditional effects Move (block1, from, to) and MoveToTable (block1, from) collapse into one Move (block1, from, to): Op(ACTION: Move(block1, from, to), PRECOND: On (block1, from) ^ Clear (block1) ^ Clear (to) EFFECT: On (block1, to) ^ Clear (from) ^ ~On(block1, from) ^ ~Clear(to) when to<>Table There’s a problem with this operator: can you spot what it is? Negated and disjunctive goals Universally quantified preconditions and effects

Reasoning About Resources Introduce numeric variables that can be used as measures These variables represent resource quantities, and change over the course of the plan Certain actions may produce (increase the quantity of) resources Other actions may consume (decrease the quantity of) resources More generally, may want different types of resources Continuous vs. discrete Sharable vs. nonsharable Reusable vs. consumable vs. self-replenishing

Other Real-World Planning Issues Conditional planning Partial observability Information gathering actions Execution monitoring and replanning Continuous planning Multi-agent (cooperative or adversarial) planning

Planning Summary Planning representations Planning approaches Situation calculus STRIPS representation: Preconditions and effects Planning approaches State-space search (STRIPS, forward chaining, ….) Plan-space search (partial-order planning, HTN, …) Constraint-based search (GraphPlan, SATplan, …) Search strategies Forward planning Goal regression Backward planning Least-commitment Nonlinear planning